The Papers

Guide

Can AI read charts and tables in a PDF? What gets lost

AI can often read simple text tables in a PDF, but complex tables lose headers and row alignment, and charts are only read where the tool supports visual retrieval — otherwise the values in a chart are effectively invisible. To keep the data, put chart values and table contents in text form with their units, periods and qualifiers, or publish a structured version alongside the PDF.

Tables: readable if they are simple

A table in a PDF is a set of text fragments placed in a grid. There is no “cell” object; the AI infers rows and columns from coordinates. Simple tables with one header row and short values usually survive. Problems start with multi-level headers, merged cells, wrapped text, totals rows, footnote markers and tables that continue on the next page.

OpenAI’s help center says ChatGPT may not reliably extract exact values from image-based tables or complex layouts, and recommends spreadsheets or text files when precision matters.

Charts: often not read at all

A chart is drawn shapes. Unless the values are printed as data labels in the text layer, there is no number to extract. OpenAI describes visual retrieval — reading charts and images inside PDFs — as available in specific contexts; in others ChatGPT uses text-only retrieval, so a chart without labels contributes nothing. Even where an image is read, values estimated from bar heights are approximations.

What gets lost, element by element

ElementWhat is lostEffect on answers
Multi-level column headersWhich year or segment a column belongs toFigures quoted for the wrong period
Merged cellsWhich rows a label applies toValues attached to the wrong item
Units in a caption (“$000s”)ScaleOff by a factor of a thousand
Footnote markers (¹, *)Qualifiers like “pro forma” or “unaudited”Estimates presented as facts
Negative numbers in parenthesesSignLosses read as gains
Chart without data labelsAll valuesMissing or guessed numbers
Legend and colorsWhich series is whichSeries swapped
Table split across pagesHeader on later pagesOrphaned rows

How to keep charts and tables readable

  • Add data labels to every chart, or a small value table under it.
  • State the takeaway of each chart in a sentence with the actual numbers.
  • Keep one header row; repeat units and periods in the header itself.
  • Avoid merged cells; repeat labels instead.
  • Write negatives with a minus sign or say “loss” in the label.
  • Keep tables on one page or repeat the header row.
  • Attach the source spreadsheet when values matter.
  • Ask the AI about one table or page at a time; OpenAI notes large, image-heavy files may not be fully analysed.

Start from the source, not the PDF

The most reliable chart data is the data the chart was built from. A PowerPoint chart stores its series values; an Excel sheet stores the cell values, number formats and formulas. Extracting from those beats reading pixels from a PDF.

The Papers reads native PowerPoint charts with their series data, tables and speaker notes, and Excel cell values with number formats, formulas and their dependencies. From a PDF it uses the text layer with positions and falls back to OCR for scanned pages. Every value in the resulting AI Version is checked against the source and keeps a reference to its page, slide or cell; anything the source cannot settle goes to you to confirm.

FAQ

Can ChatGPT read a chart image if I upload it?
In contexts where it reads images, it can describe a chart and estimate values, but estimates from bar heights are not exact. Give the numbers as text when they matter.
Why does AI mix up columns in my table?
Because the PDF has no cells, only positioned text. Multi-level headers and merged cells make the column assignment ambiguous.
Is Excel better than PDF for tables?
For exact values, yes. OpenAI recommends spreadsheets or text files when precision matters. Keep units in headers and avoid merged cells there too.
Does The Papers recreate my charts?
It extracts the data behind them — series values from native PowerPoint charts, values from tables — into the AI Version so agents read numbers rather than images.

Publish your tables and chart data as fidelity-checked text that every AI reads the same way.

Sources

Related guides